Equivalence Trial Design Principles

For example, an equivalence criterion could be expressed as:

This type of ratio-based framework is particularly familiar in bioequivalence studies.

For multiplicative effects, logarithmic transformation is often useful:

which converts a ratio problem into a difference on the log scale.

Bioequivalence vs. Clinical Equivalence

Bioequivalence and clinical equivalence are related but distinct concepts.

Bioequivalence typically concerns pharmacokinetic measures such as:

• AUC
• Cmax
• Other prespecified exposure parameters

Clinical equivalence concerns whether treatments produce sufficiently similar clinical outcomes according to a prespecified margin.

The statistical logic of confidence intervals and equivalence margins is related, but the estimands, endpoints, transformations, and regulatory requirements can differ substantially.

Estimand Considerations

An equivalence trial should define the treatment effect being estimated before the analysis begins.

Important estimand attributes include:

• Population
• Treatment conditions
• Endpoint
• Summary measure
• Handling of intercurrent events

This matters because different analysis strategies can answer different questions about equivalence.

Analysis Populations in Equivalence Trials

Equivalence trials require particularly careful consideration of analysis populations.

Common analysis sets include:

• Intention-to-treat population
• Per-protocol population
• Safety population

In superiority trials, deviations and treatment nonadherence can sometimes make treatment groups appear more similar and therefore bias against finding a difference.

In equivalence trials, that property can be problematic because reduced separation between groups can make equivalence easier to conclude.

Why Both ITT and Per-Protocol Analyses Can Matter

Because different analysis populations can have different biases in equivalence settings, investigators may prespecify complementary analyses.

For example:

The exact strategy should be prespecified and justified for the trial rather than selected after seeing which analysis supports equivalence.

Missing Data

Missing observations can have a particularly important effect in equivalence trials.

A missing-data strategy should therefore be established before database lock.

Potential approaches depend on the endpoint and estimand and may include:

• Model-based methods
• Multiple imputation
• Mixed-effects models
• Prespecified sensitivity analyses
• Pattern-mixture or tipping-point analyses

The appropriate approach depends on the mechanism and amount of missingness and the scientific estimand.

Assay Sensitivity

An important concern in equivalence and non-inferiority studies is assay sensitivity.

Assay sensitivity refers to the ability of the trial to distinguish effective treatments from ineffective treatments under the conditions of the study.

If the trial is poorly designed or the endpoint is insufficiently sensitive, two ineffective treatments could appear similar.

Therefore, demonstrating equivalence requires more than obtaining a narrow confidence interval. The study must also be scientifically capable of detecting meaningful differences if they exist.

Constancy and Historical Evidence

Historical evidence may be important when selecting equivalence margins and interpreting the reference treatment.

Investigators may need to establish that the reference treatment's historical effect remains reasonably applicable to the current trial setting.

Changes in:

• Patient population
• Background therapy
• Endpoint definitions
• Standard of care
• Comparator implementation
• Trial conduct

can affect the interpretation of historical treatment effects.

Equivalence Does Not Mean Identical

Even when equivalence is demonstrated, the estimated treatment difference is rarely exactly zero.

Suppose:

with:

and equivalence margins of ±5.

The treatment difference is not zero.

Instead, the evidence indicates that the plausible range of treatment differences is entirely within the clinically acceptable interval.

Precision Is Central to Equivalence

Consider two studies with the same observed treatment difference.

The point estimates are identical. The conclusions differ because the second study is much less precise.

This is one of the defining features of equivalence testing.

Sample Size and Precision

For a simple continuous endpoint:

As sample size increases, the confidence interval becomes narrower.

The purpose of increasing sample size in an equivalence trial is therefore not merely to increase the chance of detecting a difference. It is to obtain enough precision to rule out differences larger than the equivalence margin.

Power in an Equivalence Trial

Power is the probability of correctly demonstrating equivalence for a specified true treatment difference.

It is common to assume during planning that:

but power can also be evaluated for values away from zero.

This distinction is useful because the probability of demonstrating equivalence generally decreases as the true treatment difference approaches an equivalence boundary.

Power Depends on the True Difference

Suppose the equivalence interval is:

A study designed for 90% power when:

will not necessarily have 90% power when:

because the true effect is much closer to the upper equivalence boundary.

The actual power curve should therefore be examined when the design is important or the expected treatment difference is uncertain.

Equivalence Margin vs. Expected Treatment Difference

These two quantities should not be confused.

The equivalence margin is a clinical decision threshold. The expected treatment difference is a planning assumption.

For example:

could be the clinically acceptable difference while the design assumes:

for sample-size calculations.

The study is not trying to prove that the true difference equals zero. It is trying to demonstrate that the difference is sufficiently close to zero to remain within the acceptable region.

Equivalence and Multiple Endpoints

Some trials have multiple efficacy endpoints or multiple components that must all satisfy equivalence requirements.

This introduces an additional multiplicity problem.

For example, a study might require equivalence for:

• Primary efficacy endpoint
• Key secondary endpoint
• Another clinically important outcome

The protocol should clearly identify which endpoint establishes the primary equivalence conclusion and how multiplicity is handled.

Equivalence and Safety

Equivalence in efficacy does not imply equivalence in safety.

A treatment may demonstrate equivalent efficacy while having a different adverse-event profile.

Safety should therefore be analyzed separately using appropriate estimands, summaries, and inferential procedures.

Likewise, an equivalence conclusion for one clinical endpoint does not imply equivalence for every other outcome.

Protocol Considerations

An equivalence protocol should clearly define:

• Primary endpoint
• Treatment effect measure
• Equivalence margins
• Scientific justification for the margins
• Type I error
• Target power
• Sample-size assumptions
• Allocation ratio
• Primary analysis method
• Analysis populations
• Missing-data strategy
• Handling of protocol deviations
• Intercurrent events
• Sensitivity analyses
• Criteria for declaring equivalence

A Complete Equivalence Decision Algorithm

R Implementation: TOST for a Continuous Endpoint

For a simple two-group analysis, the TOST procedure can be implemented directly from the estimated difference and its standard error.

estimate <- 1.2
se <- 1.5

lower_margin <- -5
upper_margin <- 5

alpha <- 0.05

z <- qnorm(1 - alpha)

lower_ci <- estimate - z * se
upper_ci <- estimate + z * se

lower_ci
upper_ci

The resulting 90% confidence interval is approximately:

lower_ci
# approximately -1.27

upper_ci
# approximately 3.67

The equivalence criterion can then be evaluated directly:

equivalent <-
  lower_ci > lower_margin &&
  upper_ci < upper_margin

equivalent

The result is:

TRUE

because the complete interval lies inside:

Explicit TOST Calculations in R

The two one-sided tests can also be calculated explicitly.

estimate <- 1.2
se <- 1.5

lower_margin <- -5
upper_margin <- 5

alpha <- 0.05

# Test H01: Delta <= lower_margin
t_lower <-
  (estimate - lower_margin) / se

p_lower <-
  1 - pnorm(t_lower)

# Test H02: Delta >= upper_margin
t_upper <-
  (estimate - upper_margin) / se

p_upper <-
  pnorm(t_upper)

p_lower
p_upper

Equivalence is demonstrated when both one-sided p-values are below the prespecified significance level:

equivalent <-
  p_lower < alpha &&
  p_upper < alpha

equivalent

This gives the same inferential conclusion as the confidence-interval approach.

A Compact TOST Function

tost_equivalence <- function(
  estimate,
  se,
  lower_margin,
  upper_margin,
  alpha = 0.05
) {

  z <- qnorm(1 - alpha)

  lower_ci <-
    estimate - z * se

  upper_ci <-
    estimate + z * se

  equivalent <-
    lower_ci > lower_margin &&
    upper_ci < upper_margin

  list(
    estimate = estimate,
    lower_ci = lower_ci,
    upper_ci = upper_ci,
    equivalent = equivalent
  )
}

tost_equivalence(
  estimate = 1.2,
  se = 1.5,
  lower_margin = -5,
  upper_margin = 5
)

Power Calculation by Simulation

For more complex equivalence designs, simulation is often useful.

A basic simulation repeatedly generates treatment-group data, estimates the treatment difference, constructs the confidence interval, and records whether equivalence was demonstrated.

set.seed(123)

nsim <- 10000

n_T <- 99
n_R <- 99

mu_T <- 0
mu_R <- 0

sd <- 12

lower_margin <- -5
upper_margin <- 5

success <- logical(nsim)

for (i in seq_len(nsim)) {

  y_T <- rnorm(
    n_T,
    mean = mu_T,
    sd = sd
  )

  y_R <- rnorm(
    n_R,
    mean = mu_R,
    sd = sd
  )

  estimate <-
    mean(y_T) - mean(y_R)

  se <-
    sqrt(
      sd^2 / n_T +
      sd^2 / n_R
    )

  lower_ci <-
    estimate - qnorm(0.95) * se

  upper_ci <-
    estimate + qnorm(0.95) * se

  success[i] <-
    lower_ci > lower_margin &&
    upper_ci < upper_margin
}

mean(success)

The resulting proportion is an empirical estimate of the probability of demonstrating equivalence under the assumed true difference and variability.

Why Simulation Can Be Valuable

Closed-form sample-size equations are useful, but real equivalence studies can be substantially more complicated.

Simulation can incorporate:

• Unequal allocation
• Dropout
• Non-normal endpoints
• Repeated measurements
• Mixed-effects models
• Binary outcomes
• Time-to-event endpoints
• Covariate adjustment
• Complex missing-data mechanisms

The simulation should reproduce the planned analysis as closely as possible.

Common Equivalence Trial Design Mistakes

1. Defining equivalence as a nonsignificant superiority test. Failure to reject a difference from zero does not establish equivalence.
2. Choosing the equivalence margin after seeing the data. The margin must be justified and prespecified.
3. Using a margin that is too wide. A statistically successful trial may then demonstrate a level of similarity that is not clinically meaningful.
4. Using a margin that is too narrow without considering feasibility. The resulting sample size may become impractically large.
5. Looking only at the point estimate. Equivalence depends on the entire confidence interval.
6. Using the wrong confidence interval. For a 5% TOST analysis, the corresponding two-sided confidence interval is 90%, not 95%.
7. Ignoring variability uncertainty. Underestimating the SD can result in an underpowered study.
8. Ignoring missing data. Loss of precision and departures from the intended estimand can materially affect the equivalence conclusion.
9. Ignoring protocol deviations. In equivalence studies, deviations can bias toward apparent similarity.
10. Assuming equivalence in efficacy means equivalence in safety. Each important endpoint requires its own appropriate evaluation.
11. Confusing clinical equivalence with bioequivalence. The underlying concepts are related but the estimands, endpoints, and regulatory frameworks can differ.

What Happens When the Confidence Interval Crosses a Margin?

Suppose the equivalence margin is ±5 and the 90% confidence interval is:

The lower bound is acceptable, but the upper bound exceeds +5.

Therefore:

and equivalence is not demonstrated.

This remains true even if the point estimate is only +1.

What Happens When the Confidence Interval Is Entirely Outside?

Suppose:

The entire confidence interval is above the upper equivalence margin.

This provides evidence that the treatment difference exceeds the prespecified equivalence region in the positive direction.

The study does not demonstrate equivalence.

The interpretation of superiority depends on the endpoint direction, estimand, and prespecified statistical testing strategy.

What Happens When the Interval Is Very Wide?

Suppose:

with equivalence margins of ±5.

The interval crosses both margins.

The study has not demonstrated equivalence because the data are too imprecise to rule out clinically meaningful differences.

Equivalence Margin and Clinical Interpretation

The margin translates statistical uncertainty into a clinical decision.

For example, if a difference of five points on a validated clinical scale is considered the largest clinically acceptable difference, then:

defines the statistical boundary.

The confidence interval then answers: Given the observed data, what range of treatment differences remains plausible?

Equivalence is demonstrated when that entire plausible range remains clinically acceptable.

Equivalence and Regulatory Interpretation

Regulatory expectations for equivalence studies can vary substantially depending on the product, endpoint, indication, and development context.

A statistical equivalence calculation should therefore not be treated as a complete regulatory strategy.

The protocol and statistical analysis plan should align with the applicable regulatory guidance and the specific development program.

Sample Size Inflation for Dropout

Suppose the required evaluable sample size is:

and the anticipated dropout rate is 10%.

A simple inflation calculation is:

Thus approximately 220 participants would be enrolled to obtain approximately 198 evaluable participants under the assumed dropout rate.

The appropriate inflation method depends on the estimand, analysis method, missing-data assumptions, and whether all randomized participants contribute to the primary analysis.

Allocation Ratio

Equal allocation is often statistically efficient when treatment-group costs and variances are similar.

For unequal allocation, the variance of the treatment difference becomes:

Sample-size calculations should therefore account for the actual planned allocation ratio.

Covariate Adjustment

If the primary analysis uses a prespecified covariate-adjusted model, the sample-size calculation should reflect the expected precision of that model.

For example, baseline adjustment may reduce residual variability and improve precision.

However, investigators should not assume an arbitrary variance reduction without justification.

Equivalence With Repeated Measurements

Some clinical trials measure an outcome repeatedly over time.

In such settings, the treatment effect might be defined using:

• Change from baseline at a specified time point
• Area under the response curve
• Average treatment effect over time
• Model-based contrast

The equivalence margin must correspond to the chosen estimand.

The confidence interval should then be constructed for that specific treatment contrast.

Multiplicity and Hierarchical Testing

If multiple equivalence hypotheses must be satisfied, the statistical testing strategy should be specified in advance.

Possible structures include:

• Co-primary endpoints
• Hierarchical testing
• Multiplicity-adjusted procedures
• Gatekeeping strategies

The choice depends on the scientific objective and the interpretation of the multiple endpoints.

Interim Analyses

An equivalence trial can include interim analyses, but doing so changes the statistical design.

Repeatedly evaluating equivalence using ordinary unadjusted confidence intervals can inflate the Type I error.

If interim analyses are planned, the monitoring strategy and inferential procedure should be incorporated into the original design.

Bayesian Equivalence Approaches

Although TOST is a standard frequentist framework, Bayesian approaches can also be used to evaluate equivalence.

For example, investigators might calculate:

which is the posterior probability that the treatment difference lies inside the equivalence region.

A Bayesian decision rule might then require that this probability exceed a prespecified threshold.

Such approaches require prespecified prior distributions and decision rules.

Frequentist vs. Bayesian Interpretation

The two frameworks answer related but conceptually different statistical questions.

A Practical Equivalence Design Workflow

What Should Be Reported in the Statistical Analysis Plan?

The statistical analysis plan should make the equivalence conclusion completely reproducible.

At minimum, document:

• Primary endpoint
• Estimand
• Treatment effect measure
• Lower equivalence margin
• Upper equivalence margin
• Clinical justification for the margins
• Type I error
• Target power
• Expected treatment difference
• Variance or SD assumption
• Sample-size calculation method
• Allocation ratio
• Primary statistical model
• Confidence interval method
• TOST decision rule
• Analysis populations
• Missing-data strategy
• Protocol-deviation handling
• Sensitivity analyses
• Safety analysis

Worked Example Summary

The continuous-endpoint example can be summarized as follows.

The Most Important Concept

The most important concept in equivalence trial design is simple: equivalence must be demonstrated by ruling out clinically meaningful differences, not merely by failing to find a statistically significant difference.

The investigator first defines a clinically meaningful equivalence region:

The trial is then designed to provide enough precision to determine whether the observed treatment difference is compatible with that region.

The final statistical conclusion is based on the confidence interval:

When the entire interval lies within the margins, equivalence is demonstrated.

When the interval crosses a margin, equivalence is not demonstrated.

References

Schuirmann, D.J. (1987). A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability. Journal of Pharmacokinetics and Biopharmaceutics, 15, 657–680.
Westlake, W.J. (1976). Symmetrical confidence intervals for bioequivalence trials. Biometrics, 32, 741–744.
Westlake, W.J. (1981). Bioavailability and bioequivalence. In Topics in Pharmaceutical Sciences.
Chow, S.-C. & Liu, J.-P. (2008). Design and Analysis of Bioavailability and Bioequivalence Studies. Chapman & Hall/CRC.
Piaggio, G., Elbourne, D.R., Altman, D.G., Pocock, S.J. & Evans, S.J.W. (2006). Reporting of noninferiority and equivalence randomized trials: extension of the CONSORT statement. JAMA, 295(10), 1152–1160.
FDA. Statistical Approaches to Establishing Bioequivalence. Guidance for Industry.
FDA. Non-Inferiority Clinical Trials to Establish Effectiveness. Guidance for Industry.
EMA. Guideline on the Investigation of Bioequivalence. European Medicines Agency.